APPLICATIONS OF NEURAL NETWORKS IN CHEMISTRY .1. PREDICTION OF ELECTROPHILIC AROMATIC-SUBSTITUTION REACTIONS

A back-propagation neural network was trained on connection table representations of monosubstituted benzenes to predict the products of electrophilic aromatic substitution. Ten of 13 unknown test reactions and all 32 of the training cases were correctly predicted. With an alternative charge vector representation, 8 of 13 test cases were predicted correctly. Neural networks differ from expert systems by constructing reactivity rules implicitly from examples rather than by explicitly including rules in the expert system. The results obtained by using neural networks were comparable to those obtained from an existing chemical expert system and to predictions made by synthetic organic chemists. © 1990, American Chemical Society. All rights reserved.